Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Anthropic's Mythos Just Beat OpenAI's GPT-5.5 At Real Hacking
TL;DR
Claude recovered $400,000 in lost Bitcoin by doing patient artifact work, not magic hacking — the opening story is a guy on X who regained 5 BTC after 11 years by having Claude sift through an old college hard drive, find a pre-password-change wallet.dat, and match it with a mnemonic phrase.
Notion’s May 13 developer platform is really a bid to become the shared workspace for humans and agents — with workers, webhooks, database sync from Salesforce/Stripe/GitHub/Zendesk/Postgres, and an external agents API for Claude and Codex, Notion is making messy real company context programmable.
Anthropic’s tighter Claude limits show that agent usage breaks the old “all-you-can-eat” subscription model — after cutting off third-party use in April and then reintroducing it with monthly credit caps, Anthropic is forcing builders to treat rate limits, billing caps, and context loss as product UX issues, not back-office details.
Anthropic has likely crossed a business adoption threshold many assumed OpenAI owned — Nate says both companies are near or above $3 billion in annualized revenue, while Ramp reports Anthropic now has more verified business customers than OpenAI, even as Dario Amodei says they planned for 10x growth and got 80x.
Claude Mythos beat GPT-5.5 on independent cyber attack-chain evaluations, and that should reset security assumptions — tests from Xbow and the UK AI Security Institute showed Mythos getting farther through recon, credential theft, lateral movement, privilege escalation, and full takeover on the same token budget than GPT-5.5 or other models.
AWS managed desktops for agents makes the “we have no API” excuse much weaker — Amazon WorkSpaces now lets agents operate desktop apps with centralized logging, auditing, screenshots, and permissions, opening legacy enterprise systems like ERP, claims, finance ops, and healthcare admin to automation.
The Breakdown
The Bitcoin Wallet Story That Sets the Tone
Nate opens with a perfect example of where AI is actually useful now: not cinematic hacking, but relentless, patient research. A guy on X recovered 5 Bitcoin—about $400,000—from a wallet he’d been locked out of for 11 years by having Claude comb through old college hard-drive files, find an older wallet.dat, and line it up with a mnemonic phrase he still had.
Notion Stops Being Just a Workspace and Starts Becoming Agent Infrastructure
On May 13, Notion launched a developer platform with a CLI, hosted “workers,” webhooks, database sync, custom agent tools, and an external agents API that can bring Claude or Codex into a workspace. Nate’s point is that this is bigger than “Notion added AI” — it’s Notion trying to become the place where humans and agents share context and actually do work together.
Why This Matters for Real Company Workflows
He walks through a concrete onboarding flow: a Salesforce deal closes, a webhook triggers a Notion worker, the system spins up an onboarding workspace, pulls in notes and plan data, and an agent drafts the kickoff plan for a CSM to review. The energy here is that Notion finally gives agents access to the awkward, high-context corners of companies — the docs, rogue CRMs, customer notes, and checklists where real work starts.
Anthropic’s Claude Limits and the End of Simple AI Subscriptions
The second story is messier: Claude’s runaway success with agentic workflows has pushed Anthropic into compute trouble, and now some third-party agent usage is moving behind monthly credit caps. Nate frames this as the market learning that “all you can eat AI” means something very different when the user is an always-on coding agent instead of a human chatting occasionally.
OpenClaw, Developer Goodwill, and Why “Simple Math Wins”
Anthropic first angered developers in April by blocking third-party usage of personal Claude subscriptions, which killed side projects and pushed some users to OpenAI. Now it’s allowing some usage back with limits, but Nate says the damage is partly about messaging: developers prefer a plan where they “don’t have to do math in their head,” and OpenAI’s simpler billing posture has become a real competitive advantage.
Anthropic Is Now Neck-and-Neck With OpenAI in Revenue
Nate says the old assumption that OpenAI owned the business market no longer holds. His summary: both Anthropic and OpenAI are near or above $3 billion in annualized revenue, and Ramp’s payments data now shows Anthropic with more verified business customers than OpenAI — a sign that the race tightened fast.
Mythos vs. GPT-5.5: The Cyber Evaluations That Landed Hard
This is the biggest warning in the video. Independent testing from Xbow and the UK AI Security Institute found Claude Mythos preview outperforming GPT-5.5 and others across a serious attack chain — recon, credential theft, lateral movement, web app exploitation, privilege escalation, command-and-control persistence, infrastructure compromise, and full network takeover — while staying more token efficient.
AWS Gives Agents Desktops, and Legacy Software Is No Longer Safe From Automation
The last story sounds boring until Nate explains the implication: AI agents can now operate desktop applications inside managed Amazon WorkSpaces with logging, screenshots, permissions, and auditing. His practical caution is to start with read-only or draft mode, but the larger point is clear: if AI got this much better at computer use in just months, then the “our systems are too old” excuse is fading fast.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
